AXIS: Explainable Time Series Anomaly Detection with Large Language Models

AXIS represents a significant advancement in time-series anomaly detection (TSAD), responding to the increasing demand for explanations that clarify not only the occurrence of anomalies but also their underlying patterns. Traditional approaches using Large Language Models (LLMs) faced challenges due to their reliance on discrete tokens, making it difficult to process continuous signals effectively. AXIS overcomes this by conditioning a frozen LLM with three complementary hints: a symbolic numeric hint for grounding, a context-integrated hint for capturing dynamics, and a task-prior hint for encoding global characteristics. Extensive experiments validate AXIS's effectiveness, demonstrating its ability to yield meaningful explanations. This framework not only enhances the interpretability of anomaly detection but also sets a new benchmark for evaluating explainability in AI, making it a crucial tool for sectors that rely on accurate anomaly detection, such as finance and healthcare.